AMD has agreed to buy World Labs, the spatial intelligence startup founded by AI pioneer Fei-Fei Li, in an all-stock deal worth $8.2 billion. The acquisition gives the chipmaker its own world-model research as it tries to catch Nvidia in physical AI, the field behind robots, simulations, and autonomous machines.

[For more news, click here] 

When Fei-Fei Li appeared as a guest during AMD’s presentation at CES in January, it looked like a routine show of support from a famous AI researcher for one of her suppliers. Eight months later, AMD has agreed to buy her company. The chipmaker said on Monday it would acquire World Labs, the startup Li founded in 2024, in an all-stock deal valued at $8.2 billion.

World Labs builds world models, AI systems that learn how three-dimensional spaces are laid out and how the objects in them move, collide, and fall. Its first product, Marble, creates explorable 3D environments that are used in entertainment and to train robots in simulation. AMD already had money in the company, having joined a $1 billion funding round earlier this year.

Li, the Stanford computer scientist behind ImageNet, the dataset that helped set off the deep learning boom, will become AMD’s executive vice president and chief scientist when the deal closes, reporting directly to Chief Executive Lisa Su. Closing is expected by the end of 2026, pending regulatory approval. She has worked inside industry before, as Chief Scientist for AI at Google Cloud, and co-founded Stanford’s Institute for Human-Centered Artificial Intelligence.

Why does AMD want a world-model company?

Nvidia has spent close to two decades building software that keeps developers tied to its chips, and since January 2025 it has also offered its own open world models, called Cosmos, to robotics and self-driving car companies. AMD, on the other hand, has offered mainly text and video models to the public, which left it with little to show in one of the areas of AI where demand for computing is expected to climb sharply.

Buying World Labs gives AMD a research team that trains these models on its GPUs every day, along with an early read on what the next generation of AI workloads will demand from its hardware. AMD said the research should help guide its roadmaps across hardware, software, and systems.

In a LinkedIn post announcing the deal on Monday, Fei-Fei Li described how the two companies came to be working so closely. “We began a deep technical partnership with AMD last year, starting with model training and inference optimization on AMD GPUs,” she wrote. “As our teams worked together, we realized it would be a natural fit to bring together our AI ecosystem of software and hardware, foundation models, and applications.”

Why Physical AI Matters

Physical AI covers software that has to deal with the real world, whether that is a humanoid robot, a warehouse picker, a self-driving car, or a digital copy of a factory. A chatbot that gets a fact wrong can be corrected in its next reply. A robot arm that misjudges the weight of a box can drop it on someone’s foot, which is why developers want to train these machines in simulated worlds long before they touch anything real. Simulation also lets engineers expose a robot to thousands of rare situations, such as a spill on a warehouse floor or a pedestrian stepping out from behind a truck, that would be hard and dangerous to stage in real life.

Google DeepMind has its own Genie world models and Nvidia keeps expanding Cosmos, so a startup led by one of the best-known names in computer vision was bound to draw interest. For American manufacturers, logistics firms, and carmakers that have struggled to hire enough workers, better simulation could mean robots that are trained faster and more cheaply before they ever reach the floor.

What the World Labs deal means for AMD

The price is large for a company whose products are still young, though it is a fraction of the roughly $49 billion AMD paid for Xilinx in 2022. Under Su, AMD has been buying the pieces that sit around its chips, including the server maker ZT Systems and the Finnish AI lab Silo AI in 2024, as it tries to sell more of its Instinct accelerators against Nvidia’s data center GPUs.

AMD’s data center business has grown on the back of its AI accelerators, but Nvidia still sells the large majority of chips used to train and run AI models, and much of that lead comes from customers who do not want to rewrite software built for its platform. A world-model lab that already runs on AMD hardware gives the company a working example to show robotics and automotive customers when it asks them to switch.

AMD shares fell about 4% after the announcement, according to market reports, and because the deal is paid entirely in stock, existing shareholders will see their stakes diluted once it closes. If regulators sign off, the researcher who helped teach computers to recognize what is in a photograph will be helping decide what AMD’s next chips are built to do.

Related Articles

Pakistan’s World Context Tests Robot Training Data from Factory Workers at UC Berkeley Hackathon

Why Fortinet Is Turning to Intel to Build Its Next Cybersecurity Chip

Google Project Suncatcher Test will Send its AI Chips into Orbit